US 6978229 B1 Abstract A computer implemented method for statistical modeling and simulation of the impact of global variation and local mismatch on the performance of integrated circuits, comprises the steps of: estimating a representation of component mismatch from device performance measurements in a form suitable for circuit simulation; reducing the complexity of statistical simulation by performing a first level principal component or principal factor decomposition of global variation, including screening; further reducing the complexity of statistical simulation by performing a second level principal component decomposition including screening for each factor retained in the first level principal component decomposition step to represent local mismatch; and performing statistical simulation with the joint representation of global variation and local mismatch obtained in the second level principal component decomposition step.
Claims(6) 1. A computer implemented method for statistical modeling and simulation of the impact of global variation and local mismatch on the performance of integrated circuits, comprising the steps of:
(a) estimating a representation of component mismatch from device performance measurements in a form suitable for circuit simulation;
(b) reducing the complexity of statistical simulation by performing a first level principal component or principal factor decomposition of global variation, including screening;
(c) further reducing the complexity of statistical simulation by performing a second level principal component decomposition including screening for each factor retained in step (b) to represent local mismatch; and
(d) performing statistical simulation with the joint representation of global variation and local mismatch obtained in step (c).
2. The method of
3. The method of
4. A computer implemented method for statistical modeling and simulation of the impact of global variation and local mismatch on the performance of integrated circuits, wherein said method is integrated in a statistical design and optimization computer aided design tool to perform statistical simulation of joint and separate impact of global variation and local mismatch on performance of integrated circuits and said method comprises the steps of:
(a) estimating a representation of component mismatch from device performance measurements in a form suitable for circuit simulation;
(b) reducing the complexity of statistical simulation by performing a first level principal component or principal factor decomposition of global variation, including screening;
(c) further reducing the complexity of statistical simulation by performing a second level principal component decomposition including screening for each factor retained in step (b) to represent local mismatch; and
(d) performing statistical simulation with the joint representation of global variation and local mismatch obtained in step (c).
5. The method of
6. The method of
Description This application claims the benefit of U.S. Provisional Application Ser. No. 60/166,242 filed on Nov. 18, 1999, the contents of which is incorporated by reference as if fully set forth herein. The present invention pertains to fabrication of integrated circuits and more particularly to methods and systems for analyzing the performance and manufacturability of integrated circuits. Miniaturization and scaling of integrated circuits have resulted in two conflicting trends. On one hand, margins available in high-performance designs are shrinking, increasing the likelihood that inevitable variations during manufacturing will cause performance violations. On the other hand, both the amount of variation as a fraction of the feature sizes, and the sensitivity of the characteristics of transistors and interconnects to manufacturing variations is increasing. This makes it essential to model and simulate accurately and efficiently the impact of manufacturing variations on the performance of integrated circuits. Manufacturing variations include intra-die variations, hereinafter referred to as “local mismatch”, and inter-die variations, hereinafter referred to as “global variations”. Such intra-die and global variations are random and systematic variations in the material composition, or in the processing steps that are used to fabricate the integrated circuits. Local mismatch, or intra-die variations, are differences in the electrical properties of circuit elements that affect components of the circuits that are fabricated on the same die. For example, two identically drawn transistors fabricated next to each other in any particular chip will still present different threshold voltage values. Global variations, or inter-die variations, affect integrated circuit devices by causing random differences in the electrical properties of circuit elements, such as transistors, resistors, capacitors, memory cells, wires, etc., that are fabricated on different chips from the same wafer, on different wafers, or on different batches of wafers. As circuits are designed assuming i) a given nominal value of the electrical properties of its elements and ii) that identically drawn elements behave exactly in the same way, both global variations and local mismatch affect the actual circuit performance and are essential to determine the yield of the product. Local mismatch has not been adequately modeled in the past. Accurate matching of the electrical properties of active and passive elements is fundamental for functional and parametric performances of analog and mixed-signal integrated performance (IP) blocks, such as operational amplifiers (OPAMPs), digital to analog (D/A) and analog to digital (A/D) converters, phase locked loops (PLLs), etc. Achieving target functional and parametric yield of analog and mixed-signal components frequently represents a major bottleneck for the global time-to-volume performance of complex very large scale integrated (VLSI) systems. Previous work in modeling and analysis of matching properties of electronic devices, such as that described in the paper entitled “ First, they model the matching properties of MOSFET “macro” characteristics, such as threshold voltage (V Second, applying a device level mismatch model to the statistical simulation of electronic circuits requires the assumption that every matched device is described by a different set of low-level device parameters, each associated with a corresponding random variable (RV) (“ The variance of the relevant circuit performance parameters can then be estimated via MonteCarlo analysis. This process requires the generation of a sequence of correlated vectors of random numbers, and the evaluation of the circuit performance corresponding to each random vector instance either by directly using SPICE or via RSM macro-modeling. The dimensionality of the corresponding RV space can be very large when the simultaneous variation of all matched n-tuples of devices is considered. The problem associated with the large dimensionality of the mismatch simulation task has not yet been properly addressed. The σ-space approach of Michael and Ismail (cited above), which can be proven to be equivalent to the Choleski factorization technique used by Felt et al. (cited above), requires
The empirical approach of Guardiani et al. (cited above), has an even greater complexity, and can be only applied to very simple circuits. Conti (“ A method of modeling and simulating the impact of global and local variation on the performance of integrated circuits includes the steps of estimating a representation of component mismatch from device performance measurements in a form suitable for circuit simulation; reducing the complexity of statistical simulation by performing a first level principal component or principal factor decomposition of global variation, including screening; further reducing the complexity of statistical simulation by performing a second level principal decomposition including screening for each factor retained in the step of reducing the complexity; and performing statistical simulation with the joint representation of global and local variation obtained in the step of further reducing the complexity. A method of modeling and simulating the impact of global and local variation on the performance of integrated circuits includes the following steps. First, convert, if necessary, the device mismatch model at the device performance level into a model suitable for circuit simulation, that is, at the SPICE parameter level. Second, if the global variation is provided, perform a first level principal-component or principal-factor decomposition and screening to represent the global variation. If no global variation is provided, perform this decomposition for the local variation alone. Third, perform a second level principal-component or principal-factor decomposition and screening for each independent factor identified in the second step to represent the local variation and reduce its dimensionality. If no global variation is explained in the initial model before local variation is applied, this step does not apply. Fourth, perform statistical circuit simulation and analysis with the combined set of independent factors that result from the second and/or third steps to estimate the impact of global and local variation jointly and/or separately on the circuit of interest. In accordance with the present invention, an efficient representation of both inter-die and intra-die variation is constructed to analyze the joint impact of these sources of process variation on a design. The impact of manufacturing variations on a component is typically represented by estimating the distribution of SPICE model parameters for that component. A statistical SPICE model that represents both inter-die and intra-die variations has to account for two types of correlation; that is, correlation between model parameters and correlation between matched components. The correlation between model parameters arises because most commonly used SPICE models utilize non-independent parameters. The correlation between different components on the same die arises because of intra-die process variation. Both the inter-die and intra-die correlation can be represented in a single correlation matrix. Given n matched components (C This representation is computationally very expensive since the matrix grows as 0(nm×nm). Two simplifications are realized by using principal component (PC) decomposition. The first simplification is the use of PCS to represent the correlation between the model parameters and the second simplification is the use of PCS to represent the correlation between the matched devices. Principal component representation expresses a set of correlated random variables in terms of a set of independent random variables. This is obtained by applying a congruence transformation of the form: ΓΣΓ The correlation matrix is further simplified by using a second level of PC decomposition, replacing a PC of the original model, e.g., f, by a linear combination of a set of second-level PCS representing the intra-die variation. The result of the second PC decomposition is a unified representation that captures both the correlation between the model parameters and the matched components. Moreover, the transformation maintains the correlation between the model parameters because the RV for each original PC is replaced by an equivalent equation with the same mean and variance. The second level PC transformation adds, at worst, n new PCS, resulting in a total of n×k PCS. Since the statistical SPICE model is independent of the application, the number of PCS k is fixed, and usually k<<m, where m is the number of model parameters. This results in a mismatch simulation method that adds only 0(n) new RVs for a circuit with n matched components. An advantage of the approach of the present invention is that in practice the number of factors required is scarcely the n×k factors required in the worst case. This happens for two reasons. First, not all k PCS required to capture the correlation of the model parameters are necessary for representing mismatch. Usually, a much smaller number of variables are required. Second, because the amount of mismatch is typically very small, the correlation matrix is, in general, characterized by a small number of dominant eigenvalues. Therefore, using the equation that gives the fraction of total variance explained by the k-th PCS, the variance of the system can be well approximated by using a small number of PCS compared to the worst-case n. For example, in the asymptotic case of perfectly matched devices, this method automatically produces only one PC (corresponding to the dominant eigenvalue) for each independent process factor, resulting in no increase in the number of RVs. In the intermediate situations between the worst-case and the asymptotic case, the approach of the present invention provides an approximation scheme where a small number of PCS can be selected to accurately approximate the correlation between the components. Mismatch characterization typically does not produce the correlation between PCS of a statistical SPICE model. Usually, the result of mismatch characterization is a set of coefficients for a mismatch model of device performance. For example, MOS transistor mismatch characterization often results in the coefficients of a Pelgrom-style model for threshold voltage (V The procedure for determining the unified statistical representation starts with a statistical SPICE model for correlation between model parameters. A subset of the PCS of this model is selected for representing component mismatch. The selection is based on two considerations: the device characteristics for which mismatch has been characterized, and the weightings of the different PCS in the equation for each SPICE parameter that impact these device characteristics. For example, if mismatch characterization has been performed for threshold-voltage and gain-factor of longchannel MOSFETs, then PCS that have the most impact on the VTH0 and U0 (mobility) parameters of the BSIM3v3 model are selected. Once the PCS of the model without mismatch have been selected, a numerical optimization is performed to find the correlation between these factors in order to obtain the measured mismatch in device characteristics. Each optimization results in the correlation of the selected PC for one set of device geometry and layout parameters. Repeating this procedure for different geometries and layout distances results in a set of correlation values that are either fit with an interpolating function or represented in a look-up table. This procedure results in a model of PC correlation as a function of device geometry and layout. The step of modeling PC correlation as a function of device geometry and layout is performed only once for a particular technology and layout style. The method of the present invention can be used as part of the Circuit Surfer statistical design and verification environment. This environment supports many tasks typically required for analog and mixed signal design for manufacturability, such as statistical simulation, sensitivity analysis, response surface modeling, and circuit optimization for manufacturability. Mismatch simulation is implemented in this environment as an annotation of the circuit netlist to specify the matched components, and annotations to the statistical SPICE models to include the effect of mismatch. The annotated netlist and SPICE models are used to derive a separate statistical SPICE model for each component using the twolevel PCA. The modified netlist forms the input to Circuit Surfer. This implementation enables all the capabilities of statistical design for mismatch analyses such as variable screening, response surface modeling and MonteCarlo using mismatch factors, and optimization of a design to reduce its mismatch sensitivity. The use of the present invention is illustrated with two applications: a low voltage OPAMP and a D/A Converter. These topologies were implemented based on their continued significance in low voltage VLSI signal processing applications and their performances sensitivities to device mismatch characteristics. The following is a description of an embodiment of the method, in accordance with the present invention, for statistical modeling and simulation of the impact of global variations and local mismatch on the performance of operational amplifiers (OPAMPs) of the type described. Such local mismatch and global variations are typically caused by random and systematic variations in the material composition, or in the processing steps that are used to fabricate the integrated circuits. Global variations affect integrated circuit devices by causing random differences in the electrical properties of the circuit elements, such as transistors, resistors, capacitors, memory cells, wires, etc., that are fabricated on different chips from the same wafer, on different wafers, or on different batches of wafers. The local mismatch, or intra-die variations, are differences in the electrical properties of circuit elements that affect components of circuits that are fabricated on the same die. For example, two identically drawn transistors, fabricated next to each other in any particular chip, will still present different threshold voltage values. The OPAMP is of the 2-stage rail-to-rail class-AB architecture shown in A statistical SPICE model was estimated for the process technology used in this example. The model was estimated using measured variance and correlation between key transistor performances. This set of performances includes: transistor drive current in the linear and saturation regions (Idsat, Idlin) and transistor threshold voltage in the linear and the saturation region (Vtsat, Vtlin). The statistical SPICE model is represented in terms of principal factors, thus achieves the effect of first-level principal component decomposition. The mismatch model described above is not suitable for direct use in circuit simulation. A SPICE level representation of mismatch is derived by selecting SPICE model parameters that effect long-channel threshold voltage and the gain factor. For BSIM3v3 SPICE models, these parameters are VTH0, TOX, and U0. Numerical optimization is performed to find the correlation between these factors in order to obtain the measured mismatch in threshold voltage and gain factor. The statistical model including mismatch was used to simulate the impact of mismatch on key performance measures on this design. This simulation makes use of the method of this invention to perform the second level principle component decomposition. The impact of mismatch on this design is shown in Table 1.
A unified model is derived by applying the method described in the previous sections. The first level PCA results in nine first level PCS (pca1 . . . pca7, pcalint and pcawint) for the four matched transistors, thus resulting in a 36×36 sparsified covariance matrix. A second level PCA is performed using this matrix. Three dominant eigenvalues explain over 98% of the total variance. In this way, it is possible to account for both inter-die and intra-die variations with the addition of three variables. These three variables are suitably obtained by re-defining pca2, pca4 and pca5 as linear combinations of two independent components. Two sets of Monte-Carlo simulations were performed, one using only the inter-die model and the second using the unified model. As expected, this design shows that Vos is extremely sensitive to mismatch. By quantifying the exact impact of mismatch the method of the present invention allows an accurate assessment of manufacturability of this topology. It also shows that parametric yield estimates can be overly optimistic for designs sensitive to mismatch if mismatch effects are not accounted for in statistical simulation or worst-case models. The D/A Converter shown in Simulation with and without intra-die variation was carried out as described for the OPAMP. The procedure was deriving the statistical SPICE model using principal-factors and estimating a representation of device mismatch in form of a SPICE model are identical to those for the OPAMP. The simulation was carried out using the method of this invention. The mismatch effects were examined for the binary weighted current source identified in
Intra-die variation causes the standard deviation of DNL to increase by a factor of 13. Such a large increase in the increment causes large changes in the linearity of the D/A Converter. As in the previous example, including intra-die effects is vital to get accurate estimation of circuit performance variability, but since the circuit's performance for this topology depends upon the pairwise matching of eight different transistors, a complete statistical simulation using a standard methodology would entail running a Monte-Carlo experiment with large number of correlated RVs. For example, a circuit with eight matched transistors and 7 to 12 independent sources of process variation (typical of advanced CMOS processes) requires the generation of a sequence of correlated random numbers for as many as 96 different RVs. Several thousand SPICE evaluations would be required to stabilize the Monte-Carlo results for such a high-dimensional system. The advantages of complexity reduction using second order principal component decomposition are illustrated by the graph depicted in Patent Citations
Non-Patent Citations
Referenced by
Classifications
Legal Events
Rotate |